LiteLLM Agent Platform
A tiernew this weekSelf-hosted Kubernetes infrastructure for running coding agents like Claude Code and Codex in isolated, persistent sandboxes — built for teams who want managed agent infra without giving up data control.
Kai's verdict
A genuinely useful piece of missing infrastructure for teams already running LiteLLM Gateway who want to graduate from 'agent in a script' to 'agent in production' — but you'll need real Kubernetes chops and tolerance for alpha-stage rough edges before this earns a place in a critical workflow. (Verdict pending Phi's full review.)
Strengths
- Full data residency — nothing leaves your own Kubernetes cluster, making it viable for regulated industries
- Persistent session state across pod restarts via Postgres, solving the #1 pain point of stateful agents in production
- Vault proxy for secrets: sandbox containers get stub credentials; real keys are injected on outbound TLS, not baked into pods
- Plugs directly into LiteLLM Gateway for model routing, cost tracking, and rate limiting across 100+ providers
- Quickstart is genuinely fast — two shell commands stand up a local kind cluster with working sandboxes
Weaknesses
- Alpha-quality: rough edges, community-only support, no SLA — not for teams that need production guarantees today
- Meaningful DevOps overhead: requires Docker, Kubernetes (kind or EKS), Postgres, and Helm familiarity just to get started
- Sessions expire after 24h; no longer-lived persistence or cross-session memory story yet
Best for
Platform and DevOps engineers at orgs that need to run multi-agent coding workflows (Claude Code, Codex) in production, with full infrastructure ownership and strict data residency requirements.
Pricing
Free (open source, MIT)
Free to self-host; you pay your own cloud infrastructure costs (AWS EKS, Render, etc.). No SaaS tier — fully bring-your-own-infra.